Fix "delayed_work_pending()" macro expansion
[sfrench/cifs-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/hash.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15
16 /*
17  * tunables
18  */
19 static const int cfq_quantum = 4;               /* max queue in one round of service */
20 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
21 static const int cfq_back_max = 16 * 1024;      /* maximum backwards seek, in KiB */
22 static const int cfq_back_penalty = 2;          /* penalty of a backwards seek */
23
24 static const int cfq_slice_sync = HZ / 10;
25 static int cfq_slice_async = HZ / 25;
26 static const int cfq_slice_async_rq = 2;
27 static int cfq_slice_idle = HZ / 125;
28
29 #define CFQ_IDLE_GRACE          (HZ / 10)
30 #define CFQ_SLICE_SCALE         (5)
31
32 #define CFQ_KEY_ASYNC           (0)
33
34 /*
35  * for the hash of cfqq inside the cfqd
36  */
37 #define CFQ_QHASH_SHIFT         6
38 #define CFQ_QHASH_ENTRIES       (1 << CFQ_QHASH_SHIFT)
39 #define list_entry_qhash(entry) hlist_entry((entry), struct cfq_queue, cfq_hash)
40
41 #define list_entry_cfqq(ptr)    list_entry((ptr), struct cfq_queue, cfq_list)
42
43 #define RQ_CIC(rq)              ((struct cfq_io_context*)(rq)->elevator_private)
44 #define RQ_CFQQ(rq)             ((rq)->elevator_private2)
45
46 static struct kmem_cache *cfq_pool;
47 static struct kmem_cache *cfq_ioc_pool;
48
49 static DEFINE_PER_CPU(unsigned long, ioc_count);
50 static struct completion *ioc_gone;
51
52 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
53 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
54 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
55
56 #define ASYNC                   (0)
57 #define SYNC                    (1)
58
59 #define cfq_cfqq_dispatched(cfqq)       \
60         ((cfqq)->on_dispatch[ASYNC] + (cfqq)->on_dispatch[SYNC])
61
62 #define cfq_cfqq_class_sync(cfqq)       ((cfqq)->key != CFQ_KEY_ASYNC)
63
64 #define cfq_cfqq_sync(cfqq)             \
65         (cfq_cfqq_class_sync(cfqq) || (cfqq)->on_dispatch[SYNC])
66
67 #define sample_valid(samples)   ((samples) > 80)
68
69 /*
70  * Per block device queue structure
71  */
72 struct cfq_data {
73         request_queue_t *queue;
74
75         /*
76          * rr list of queues with requests and the count of them
77          */
78         struct list_head rr_list[CFQ_PRIO_LISTS];
79         struct list_head busy_rr;
80         struct list_head cur_rr;
81         struct list_head idle_rr;
82         unsigned int busy_queues;
83
84         /*
85          * cfqq lookup hash
86          */
87         struct hlist_head *cfq_hash;
88
89         int rq_in_driver;
90         int hw_tag;
91
92         /*
93          * idle window management
94          */
95         struct timer_list idle_slice_timer;
96         struct work_struct unplug_work;
97
98         struct cfq_queue *active_queue;
99         struct cfq_io_context *active_cic;
100         int cur_prio, cur_end_prio;
101         unsigned int dispatch_slice;
102
103         struct timer_list idle_class_timer;
104
105         sector_t last_sector;
106         unsigned long last_end_request;
107
108         /*
109          * tunables, see top of file
110          */
111         unsigned int cfq_quantum;
112         unsigned int cfq_fifo_expire[2];
113         unsigned int cfq_back_penalty;
114         unsigned int cfq_back_max;
115         unsigned int cfq_slice[2];
116         unsigned int cfq_slice_async_rq;
117         unsigned int cfq_slice_idle;
118
119         struct list_head cic_list;
120 };
121
122 /*
123  * Per process-grouping structure
124  */
125 struct cfq_queue {
126         /* reference count */
127         atomic_t ref;
128         /* parent cfq_data */
129         struct cfq_data *cfqd;
130         /* cfqq lookup hash */
131         struct hlist_node cfq_hash;
132         /* hash key */
133         unsigned int key;
134         /* member of the rr/busy/cur/idle cfqd list */
135         struct list_head cfq_list;
136         /* sorted list of pending requests */
137         struct rb_root sort_list;
138         /* if fifo isn't expired, next request to serve */
139         struct request *next_rq;
140         /* requests queued in sort_list */
141         int queued[2];
142         /* currently allocated requests */
143         int allocated[2];
144         /* pending metadata requests */
145         int meta_pending;
146         /* fifo list of requests in sort_list */
147         struct list_head fifo;
148
149         unsigned long slice_start;
150         unsigned long slice_end;
151         unsigned long slice_left;
152
153         /* number of requests that are on the dispatch list */
154         int on_dispatch[2];
155
156         /* io prio of this group */
157         unsigned short ioprio, org_ioprio;
158         unsigned short ioprio_class, org_ioprio_class;
159
160         /* various state flags, see below */
161         unsigned int flags;
162 };
163
164 enum cfqq_state_flags {
165         CFQ_CFQQ_FLAG_on_rr = 0,
166         CFQ_CFQQ_FLAG_wait_request,
167         CFQ_CFQQ_FLAG_must_alloc,
168         CFQ_CFQQ_FLAG_must_alloc_slice,
169         CFQ_CFQQ_FLAG_must_dispatch,
170         CFQ_CFQQ_FLAG_fifo_expire,
171         CFQ_CFQQ_FLAG_idle_window,
172         CFQ_CFQQ_FLAG_prio_changed,
173         CFQ_CFQQ_FLAG_queue_new,
174 };
175
176 #define CFQ_CFQQ_FNS(name)                                              \
177 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
178 {                                                                       \
179         cfqq->flags |= (1 << CFQ_CFQQ_FLAG_##name);                     \
180 }                                                                       \
181 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
182 {                                                                       \
183         cfqq->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                    \
184 }                                                                       \
185 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
186 {                                                                       \
187         return (cfqq->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;        \
188 }
189
190 CFQ_CFQQ_FNS(on_rr);
191 CFQ_CFQQ_FNS(wait_request);
192 CFQ_CFQQ_FNS(must_alloc);
193 CFQ_CFQQ_FNS(must_alloc_slice);
194 CFQ_CFQQ_FNS(must_dispatch);
195 CFQ_CFQQ_FNS(fifo_expire);
196 CFQ_CFQQ_FNS(idle_window);
197 CFQ_CFQQ_FNS(prio_changed);
198 CFQ_CFQQ_FNS(queue_new);
199 #undef CFQ_CFQQ_FNS
200
201 static struct cfq_queue *cfq_find_cfq_hash(struct cfq_data *, unsigned int, unsigned short);
202 static void cfq_dispatch_insert(request_queue_t *, struct request *);
203 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk, gfp_t gfp_mask);
204
205 /*
206  * scheduler run of queue, if there are requests pending and no one in the
207  * driver that will restart queueing
208  */
209 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
210 {
211         if (cfqd->busy_queues)
212                 kblockd_schedule_work(&cfqd->unplug_work);
213 }
214
215 static int cfq_queue_empty(request_queue_t *q)
216 {
217         struct cfq_data *cfqd = q->elevator->elevator_data;
218
219         return !cfqd->busy_queues;
220 }
221
222 static inline pid_t cfq_queue_pid(struct task_struct *task, int rw, int is_sync)
223 {
224         /*
225          * Use the per-process queue, for read requests and syncronous writes
226          */
227         if (!(rw & REQ_RW) || is_sync)
228                 return task->pid;
229
230         return CFQ_KEY_ASYNC;
231 }
232
233 /*
234  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
235  * We choose the request that is closest to the head right now. Distance
236  * behind the head is penalized and only allowed to a certain extent.
237  */
238 static struct request *
239 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
240 {
241         sector_t last, s1, s2, d1 = 0, d2 = 0;
242         unsigned long back_max;
243 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
244 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
245         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
246
247         if (rq1 == NULL || rq1 == rq2)
248                 return rq2;
249         if (rq2 == NULL)
250                 return rq1;
251
252         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
253                 return rq1;
254         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
255                 return rq2;
256         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
257                 return rq1;
258         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
259                 return rq2;
260
261         s1 = rq1->sector;
262         s2 = rq2->sector;
263
264         last = cfqd->last_sector;
265
266         /*
267          * by definition, 1KiB is 2 sectors
268          */
269         back_max = cfqd->cfq_back_max * 2;
270
271         /*
272          * Strict one way elevator _except_ in the case where we allow
273          * short backward seeks which are biased as twice the cost of a
274          * similar forward seek.
275          */
276         if (s1 >= last)
277                 d1 = s1 - last;
278         else if (s1 + back_max >= last)
279                 d1 = (last - s1) * cfqd->cfq_back_penalty;
280         else
281                 wrap |= CFQ_RQ1_WRAP;
282
283         if (s2 >= last)
284                 d2 = s2 - last;
285         else if (s2 + back_max >= last)
286                 d2 = (last - s2) * cfqd->cfq_back_penalty;
287         else
288                 wrap |= CFQ_RQ2_WRAP;
289
290         /* Found required data */
291
292         /*
293          * By doing switch() on the bit mask "wrap" we avoid having to
294          * check two variables for all permutations: --> faster!
295          */
296         switch (wrap) {
297         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
298                 if (d1 < d2)
299                         return rq1;
300                 else if (d2 < d1)
301                         return rq2;
302                 else {
303                         if (s1 >= s2)
304                                 return rq1;
305                         else
306                                 return rq2;
307                 }
308
309         case CFQ_RQ2_WRAP:
310                 return rq1;
311         case CFQ_RQ1_WRAP:
312                 return rq2;
313         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
314         default:
315                 /*
316                  * Since both rqs are wrapped,
317                  * start with the one that's further behind head
318                  * (--> only *one* back seek required),
319                  * since back seek takes more time than forward.
320                  */
321                 if (s1 <= s2)
322                         return rq1;
323                 else
324                         return rq2;
325         }
326 }
327
328 /*
329  * would be nice to take fifo expire time into account as well
330  */
331 static struct request *
332 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
333                   struct request *last)
334 {
335         struct rb_node *rbnext = rb_next(&last->rb_node);
336         struct rb_node *rbprev = rb_prev(&last->rb_node);
337         struct request *next = NULL, *prev = NULL;
338
339         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
340
341         if (rbprev)
342                 prev = rb_entry_rq(rbprev);
343
344         if (rbnext)
345                 next = rb_entry_rq(rbnext);
346         else {
347                 rbnext = rb_first(&cfqq->sort_list);
348                 if (rbnext && rbnext != &last->rb_node)
349                         next = rb_entry_rq(rbnext);
350         }
351
352         return cfq_choose_req(cfqd, next, prev);
353 }
354
355 static void cfq_resort_rr_list(struct cfq_queue *cfqq, int preempted)
356 {
357         struct cfq_data *cfqd = cfqq->cfqd;
358         struct list_head *list;
359
360         BUG_ON(!cfq_cfqq_on_rr(cfqq));
361
362         list_del(&cfqq->cfq_list);
363
364         if (cfq_class_rt(cfqq))
365                 list = &cfqd->cur_rr;
366         else if (cfq_class_idle(cfqq))
367                 list = &cfqd->idle_rr;
368         else {
369                 /*
370                  * if cfqq has requests in flight, don't allow it to be
371                  * found in cfq_set_active_queue before it has finished them.
372                  * this is done to increase fairness between a process that
373                  * has lots of io pending vs one that only generates one
374                  * sporadically or synchronously
375                  */
376                 if (cfq_cfqq_dispatched(cfqq))
377                         list = &cfqd->busy_rr;
378                 else
379                         list = &cfqd->rr_list[cfqq->ioprio];
380         }
381
382         /*
383          * If this queue was preempted or is new (never been serviced), let
384          * it be added first for fairness but beind other new queues.
385          * Otherwise, just add to the back  of the list.
386          */
387         if (preempted || cfq_cfqq_queue_new(cfqq)) {
388                 struct list_head *n = list;
389                 struct cfq_queue *__cfqq;
390
391                 while (n->next != list) {
392                         __cfqq = list_entry_cfqq(n->next);
393                         if (!cfq_cfqq_queue_new(__cfqq))
394                                 break;
395
396                         n = n->next;
397                 }
398
399                 list = n;
400         }
401
402         list_add_tail(&cfqq->cfq_list, list);
403 }
404
405 /*
406  * add to busy list of queues for service, trying to be fair in ordering
407  * the pending list according to last request service
408  */
409 static inline void
410 cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
411 {
412         BUG_ON(cfq_cfqq_on_rr(cfqq));
413         cfq_mark_cfqq_on_rr(cfqq);
414         cfqd->busy_queues++;
415
416         cfq_resort_rr_list(cfqq, 0);
417 }
418
419 static inline void
420 cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
421 {
422         BUG_ON(!cfq_cfqq_on_rr(cfqq));
423         cfq_clear_cfqq_on_rr(cfqq);
424         list_del_init(&cfqq->cfq_list);
425
426         BUG_ON(!cfqd->busy_queues);
427         cfqd->busy_queues--;
428 }
429
430 /*
431  * rb tree support functions
432  */
433 static inline void cfq_del_rq_rb(struct request *rq)
434 {
435         struct cfq_queue *cfqq = RQ_CFQQ(rq);
436         struct cfq_data *cfqd = cfqq->cfqd;
437         const int sync = rq_is_sync(rq);
438
439         BUG_ON(!cfqq->queued[sync]);
440         cfqq->queued[sync]--;
441
442         elv_rb_del(&cfqq->sort_list, rq);
443
444         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
445                 cfq_del_cfqq_rr(cfqd, cfqq);
446 }
447
448 static void cfq_add_rq_rb(struct request *rq)
449 {
450         struct cfq_queue *cfqq = RQ_CFQQ(rq);
451         struct cfq_data *cfqd = cfqq->cfqd;
452         struct request *__alias;
453
454         cfqq->queued[rq_is_sync(rq)]++;
455
456         /*
457          * looks a little odd, but the first insert might return an alias.
458          * if that happens, put the alias on the dispatch list
459          */
460         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
461                 cfq_dispatch_insert(cfqd->queue, __alias);
462
463         if (!cfq_cfqq_on_rr(cfqq))
464                 cfq_add_cfqq_rr(cfqd, cfqq);
465 }
466
467 static inline void
468 cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
469 {
470         elv_rb_del(&cfqq->sort_list, rq);
471         cfqq->queued[rq_is_sync(rq)]--;
472         cfq_add_rq_rb(rq);
473 }
474
475 static struct request *
476 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
477 {
478         struct task_struct *tsk = current;
479         pid_t key = cfq_queue_pid(tsk, bio_data_dir(bio), bio_sync(bio));
480         struct cfq_queue *cfqq;
481
482         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
483         if (cfqq) {
484                 sector_t sector = bio->bi_sector + bio_sectors(bio);
485
486                 return elv_rb_find(&cfqq->sort_list, sector);
487         }
488
489         return NULL;
490 }
491
492 static void cfq_activate_request(request_queue_t *q, struct request *rq)
493 {
494         struct cfq_data *cfqd = q->elevator->elevator_data;
495
496         cfqd->rq_in_driver++;
497
498         /*
499          * If the depth is larger 1, it really could be queueing. But lets
500          * make the mark a little higher - idling could still be good for
501          * low queueing, and a low queueing number could also just indicate
502          * a SCSI mid layer like behaviour where limit+1 is often seen.
503          */
504         if (!cfqd->hw_tag && cfqd->rq_in_driver > 4)
505                 cfqd->hw_tag = 1;
506 }
507
508 static void cfq_deactivate_request(request_queue_t *q, struct request *rq)
509 {
510         struct cfq_data *cfqd = q->elevator->elevator_data;
511
512         WARN_ON(!cfqd->rq_in_driver);
513         cfqd->rq_in_driver--;
514 }
515
516 static void cfq_remove_request(struct request *rq)
517 {
518         struct cfq_queue *cfqq = RQ_CFQQ(rq);
519
520         if (cfqq->next_rq == rq)
521                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
522
523         list_del_init(&rq->queuelist);
524         cfq_del_rq_rb(rq);
525
526         if (rq_is_meta(rq)) {
527                 WARN_ON(!cfqq->meta_pending);
528                 cfqq->meta_pending--;
529         }
530 }
531
532 static int
533 cfq_merge(request_queue_t *q, struct request **req, struct bio *bio)
534 {
535         struct cfq_data *cfqd = q->elevator->elevator_data;
536         struct request *__rq;
537
538         __rq = cfq_find_rq_fmerge(cfqd, bio);
539         if (__rq && elv_rq_merge_ok(__rq, bio)) {
540                 *req = __rq;
541                 return ELEVATOR_FRONT_MERGE;
542         }
543
544         return ELEVATOR_NO_MERGE;
545 }
546
547 static void cfq_merged_request(request_queue_t *q, struct request *req,
548                                int type)
549 {
550         if (type == ELEVATOR_FRONT_MERGE) {
551                 struct cfq_queue *cfqq = RQ_CFQQ(req);
552
553                 cfq_reposition_rq_rb(cfqq, req);
554         }
555 }
556
557 static void
558 cfq_merged_requests(request_queue_t *q, struct request *rq,
559                     struct request *next)
560 {
561         /*
562          * reposition in fifo if next is older than rq
563          */
564         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
565             time_before(next->start_time, rq->start_time))
566                 list_move(&rq->queuelist, &next->queuelist);
567
568         cfq_remove_request(next);
569 }
570
571 static inline void
572 __cfq_set_active_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
573 {
574         if (cfqq) {
575                 /*
576                  * stop potential idle class queues waiting service
577                  */
578                 del_timer(&cfqd->idle_class_timer);
579
580                 cfqq->slice_start = jiffies;
581                 cfqq->slice_end = 0;
582                 cfqq->slice_left = 0;
583                 cfq_clear_cfqq_must_alloc_slice(cfqq);
584                 cfq_clear_cfqq_fifo_expire(cfqq);
585         }
586
587         cfqd->active_queue = cfqq;
588 }
589
590 /*
591  * current cfqq expired its slice (or was too idle), select new one
592  */
593 static void
594 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
595                     int preempted)
596 {
597         unsigned long now = jiffies;
598
599         if (cfq_cfqq_wait_request(cfqq))
600                 del_timer(&cfqd->idle_slice_timer);
601
602         if (!preempted && !cfq_cfqq_dispatched(cfqq))
603                 cfq_schedule_dispatch(cfqd);
604
605         cfq_clear_cfqq_must_dispatch(cfqq);
606         cfq_clear_cfqq_wait_request(cfqq);
607         cfq_clear_cfqq_queue_new(cfqq);
608
609         /*
610          * store what was left of this slice, if the queue idled out
611          * or was preempted
612          */
613         if (time_after(cfqq->slice_end, now))
614                 cfqq->slice_left = cfqq->slice_end - now;
615         else
616                 cfqq->slice_left = 0;
617
618         if (cfq_cfqq_on_rr(cfqq))
619                 cfq_resort_rr_list(cfqq, preempted);
620
621         if (cfqq == cfqd->active_queue)
622                 cfqd->active_queue = NULL;
623
624         if (cfqd->active_cic) {
625                 put_io_context(cfqd->active_cic->ioc);
626                 cfqd->active_cic = NULL;
627         }
628
629         cfqd->dispatch_slice = 0;
630 }
631
632 static inline void cfq_slice_expired(struct cfq_data *cfqd, int preempted)
633 {
634         struct cfq_queue *cfqq = cfqd->active_queue;
635
636         if (cfqq)
637                 __cfq_slice_expired(cfqd, cfqq, preempted);
638 }
639
640 /*
641  * 0
642  * 0,1
643  * 0,1,2
644  * 0,1,2,3
645  * 0,1,2,3,4
646  * 0,1,2,3,4,5
647  * 0,1,2,3,4,5,6
648  * 0,1,2,3,4,5,6,7
649  */
650 static int cfq_get_next_prio_level(struct cfq_data *cfqd)
651 {
652         int prio, wrap;
653
654         prio = -1;
655         wrap = 0;
656         do {
657                 int p;
658
659                 for (p = cfqd->cur_prio; p <= cfqd->cur_end_prio; p++) {
660                         if (!list_empty(&cfqd->rr_list[p])) {
661                                 prio = p;
662                                 break;
663                         }
664                 }
665
666                 if (prio != -1)
667                         break;
668                 cfqd->cur_prio = 0;
669                 if (++cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
670                         cfqd->cur_end_prio = 0;
671                         if (wrap)
672                                 break;
673                         wrap = 1;
674                 }
675         } while (1);
676
677         if (unlikely(prio == -1))
678                 return -1;
679
680         BUG_ON(prio >= CFQ_PRIO_LISTS);
681
682         list_splice_init(&cfqd->rr_list[prio], &cfqd->cur_rr);
683
684         cfqd->cur_prio = prio + 1;
685         if (cfqd->cur_prio > cfqd->cur_end_prio) {
686                 cfqd->cur_end_prio = cfqd->cur_prio;
687                 cfqd->cur_prio = 0;
688         }
689         if (cfqd->cur_end_prio == CFQ_PRIO_LISTS) {
690                 cfqd->cur_prio = 0;
691                 cfqd->cur_end_prio = 0;
692         }
693
694         return prio;
695 }
696
697 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
698 {
699         struct cfq_queue *cfqq = NULL;
700
701         if (!list_empty(&cfqd->cur_rr) || cfq_get_next_prio_level(cfqd) != -1) {
702                 /*
703                  * if current list is non-empty, grab first entry. if it is
704                  * empty, get next prio level and grab first entry then if any
705                  * are spliced
706                  */
707                 cfqq = list_entry_cfqq(cfqd->cur_rr.next);
708         } else if (!list_empty(&cfqd->busy_rr)) {
709                 /*
710                  * If no new queues are available, check if the busy list has
711                  * some before falling back to idle io.
712                  */
713                 cfqq = list_entry_cfqq(cfqd->busy_rr.next);
714         } else if (!list_empty(&cfqd->idle_rr)) {
715                 /*
716                  * if we have idle queues and no rt or be queues had pending
717                  * requests, either allow immediate service if the grace period
718                  * has passed or arm the idle grace timer
719                  */
720                 unsigned long end = cfqd->last_end_request + CFQ_IDLE_GRACE;
721
722                 if (time_after_eq(jiffies, end))
723                         cfqq = list_entry_cfqq(cfqd->idle_rr.next);
724                 else
725                         mod_timer(&cfqd->idle_class_timer, end);
726         }
727
728         __cfq_set_active_queue(cfqd, cfqq);
729         return cfqq;
730 }
731
732 #define CIC_SEEKY(cic) ((cic)->seek_mean > (128 * 1024))
733
734 static int cfq_arm_slice_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
735
736 {
737         struct cfq_io_context *cic;
738         unsigned long sl;
739
740         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
741         WARN_ON(cfqq != cfqd->active_queue);
742
743         /*
744          * idle is disabled, either manually or by past process history
745          */
746         if (!cfqd->cfq_slice_idle)
747                 return 0;
748         if (!cfq_cfqq_idle_window(cfqq))
749                 return 0;
750         /*
751          * task has exited, don't wait
752          */
753         cic = cfqd->active_cic;
754         if (!cic || !cic->ioc->task)
755                 return 0;
756
757         cfq_mark_cfqq_must_dispatch(cfqq);
758         cfq_mark_cfqq_wait_request(cfqq);
759
760         sl = min(cfqq->slice_end - 1, (unsigned long) cfqd->cfq_slice_idle);
761
762         /*
763          * we don't want to idle for seeks, but we do want to allow
764          * fair distribution of slice time for a process doing back-to-back
765          * seeks. so allow a little bit of time for him to submit a new rq
766          */
767         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
768                 sl = min(sl, msecs_to_jiffies(2));
769
770         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
771         return 1;
772 }
773
774 static void cfq_dispatch_insert(request_queue_t *q, struct request *rq)
775 {
776         struct cfq_data *cfqd = q->elevator->elevator_data;
777         struct cfq_queue *cfqq = RQ_CFQQ(rq);
778
779         cfq_remove_request(rq);
780         cfqq->on_dispatch[rq_is_sync(rq)]++;
781         elv_dispatch_sort(q, rq);
782
783         rq = list_entry(q->queue_head.prev, struct request, queuelist);
784         cfqd->last_sector = rq->sector + rq->nr_sectors;
785 }
786
787 /*
788  * return expired entry, or NULL to just start from scratch in rbtree
789  */
790 static inline struct request *cfq_check_fifo(struct cfq_queue *cfqq)
791 {
792         struct cfq_data *cfqd = cfqq->cfqd;
793         struct request *rq;
794         int fifo;
795
796         if (cfq_cfqq_fifo_expire(cfqq))
797                 return NULL;
798         if (list_empty(&cfqq->fifo))
799                 return NULL;
800
801         fifo = cfq_cfqq_class_sync(cfqq);
802         rq = rq_entry_fifo(cfqq->fifo.next);
803
804         if (time_after(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo])) {
805                 cfq_mark_cfqq_fifo_expire(cfqq);
806                 return rq;
807         }
808
809         return NULL;
810 }
811
812 /*
813  * Scale schedule slice based on io priority. Use the sync time slice only
814  * if a queue is marked sync and has sync io queued. A sync queue with async
815  * io only, should not get full sync slice length.
816  */
817 static inline int
818 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
819 {
820         const int base_slice = cfqd->cfq_slice[cfq_cfqq_sync(cfqq)];
821
822         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
823
824         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - cfqq->ioprio));
825 }
826
827 static inline void
828 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
829 {
830         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
831 }
832
833 static inline int
834 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
835 {
836         const int base_rq = cfqd->cfq_slice_async_rq;
837
838         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
839
840         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
841 }
842
843 /*
844  * get next queue for service
845  */
846 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
847 {
848         unsigned long now = jiffies;
849         struct cfq_queue *cfqq;
850
851         cfqq = cfqd->active_queue;
852         if (!cfqq)
853                 goto new_queue;
854
855         /*
856          * slice has expired
857          */
858         if (!cfq_cfqq_must_dispatch(cfqq) && time_after(now, cfqq->slice_end))
859                 goto expire;
860
861         /*
862          * if queue has requests, dispatch one. if not, check if
863          * enough slice is left to wait for one
864          */
865         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
866                 goto keep_queue;
867         else if (cfq_cfqq_dispatched(cfqq)) {
868                 cfqq = NULL;
869                 goto keep_queue;
870         } else if (cfq_cfqq_class_sync(cfqq)) {
871                 if (cfq_arm_slice_timer(cfqd, cfqq))
872                         return NULL;
873         }
874
875 expire:
876         cfq_slice_expired(cfqd, 0);
877 new_queue:
878         cfqq = cfq_set_active_queue(cfqd);
879 keep_queue:
880         return cfqq;
881 }
882
883 static int
884 __cfq_dispatch_requests(struct cfq_data *cfqd, struct cfq_queue *cfqq,
885                         int max_dispatch)
886 {
887         int dispatched = 0;
888
889         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
890
891         do {
892                 struct request *rq;
893
894                 /*
895                  * follow expired path, else get first next available
896                  */
897                 if ((rq = cfq_check_fifo(cfqq)) == NULL)
898                         rq = cfqq->next_rq;
899
900                 /*
901                  * finally, insert request into driver dispatch list
902                  */
903                 cfq_dispatch_insert(cfqd->queue, rq);
904
905                 cfqd->dispatch_slice++;
906                 dispatched++;
907
908                 if (!cfqd->active_cic) {
909                         atomic_inc(&RQ_CIC(rq)->ioc->refcount);
910                         cfqd->active_cic = RQ_CIC(rq);
911                 }
912
913                 if (RB_EMPTY_ROOT(&cfqq->sort_list))
914                         break;
915
916         } while (dispatched < max_dispatch);
917
918         /*
919          * if slice end isn't set yet, set it.
920          */
921         if (!cfqq->slice_end)
922                 cfq_set_prio_slice(cfqd, cfqq);
923
924         /*
925          * expire an async queue immediately if it has used up its slice. idle
926          * queue always expire after 1 dispatch round.
927          */
928         if ((!cfq_cfqq_sync(cfqq) &&
929             cfqd->dispatch_slice >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
930             cfq_class_idle(cfqq) ||
931             !cfq_cfqq_idle_window(cfqq))
932                 cfq_slice_expired(cfqd, 0);
933
934         return dispatched;
935 }
936
937 static int
938 cfq_forced_dispatch_cfqqs(struct list_head *list)
939 {
940         struct cfq_queue *cfqq, *next;
941         int dispatched;
942
943         dispatched = 0;
944         list_for_each_entry_safe(cfqq, next, list, cfq_list) {
945                 while (cfqq->next_rq) {
946                         cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
947                         dispatched++;
948                 }
949                 BUG_ON(!list_empty(&cfqq->fifo));
950         }
951
952         return dispatched;
953 }
954
955 static int
956 cfq_forced_dispatch(struct cfq_data *cfqd)
957 {
958         int i, dispatched = 0;
959
960         for (i = 0; i < CFQ_PRIO_LISTS; i++)
961                 dispatched += cfq_forced_dispatch_cfqqs(&cfqd->rr_list[i]);
962
963         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->busy_rr);
964         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->cur_rr);
965         dispatched += cfq_forced_dispatch_cfqqs(&cfqd->idle_rr);
966
967         cfq_slice_expired(cfqd, 0);
968
969         BUG_ON(cfqd->busy_queues);
970
971         return dispatched;
972 }
973
974 static int
975 cfq_dispatch_requests(request_queue_t *q, int force)
976 {
977         struct cfq_data *cfqd = q->elevator->elevator_data;
978         struct cfq_queue *cfqq, *prev_cfqq;
979         int dispatched;
980
981         if (!cfqd->busy_queues)
982                 return 0;
983
984         if (unlikely(force))
985                 return cfq_forced_dispatch(cfqd);
986
987         dispatched = 0;
988         prev_cfqq = NULL;
989         while ((cfqq = cfq_select_queue(cfqd)) != NULL) {
990                 int max_dispatch;
991
992                 /*
993                  * Don't repeat dispatch from the previous queue.
994                  */
995                 if (prev_cfqq == cfqq)
996                         break;
997
998                 cfq_clear_cfqq_must_dispatch(cfqq);
999                 cfq_clear_cfqq_wait_request(cfqq);
1000                 del_timer(&cfqd->idle_slice_timer);
1001
1002                 max_dispatch = cfqd->cfq_quantum;
1003                 if (cfq_class_idle(cfqq))
1004                         max_dispatch = 1;
1005
1006                 dispatched += __cfq_dispatch_requests(cfqd, cfqq, max_dispatch);
1007
1008                 /*
1009                  * If the dispatch cfqq has idling enabled and is still
1010                  * the active queue, break out.
1011                  */
1012                 if (cfq_cfqq_idle_window(cfqq) && cfqd->active_queue)
1013                         break;
1014
1015                 prev_cfqq = cfqq;
1016         }
1017
1018         return dispatched;
1019 }
1020
1021 /*
1022  * task holds one reference to the queue, dropped when task exits. each rq
1023  * in-flight on this queue also holds a reference, dropped when rq is freed.
1024  *
1025  * queue lock must be held here.
1026  */
1027 static void cfq_put_queue(struct cfq_queue *cfqq)
1028 {
1029         struct cfq_data *cfqd = cfqq->cfqd;
1030
1031         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1032
1033         if (!atomic_dec_and_test(&cfqq->ref))
1034                 return;
1035
1036         BUG_ON(rb_first(&cfqq->sort_list));
1037         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1038         BUG_ON(cfq_cfqq_on_rr(cfqq));
1039
1040         if (unlikely(cfqd->active_queue == cfqq))
1041                 __cfq_slice_expired(cfqd, cfqq, 0);
1042
1043         /*
1044          * it's on the empty list and still hashed
1045          */
1046         list_del(&cfqq->cfq_list);
1047         hlist_del(&cfqq->cfq_hash);
1048         kmem_cache_free(cfq_pool, cfqq);
1049 }
1050
1051 static struct cfq_queue *
1052 __cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned int prio,
1053                     const int hashval)
1054 {
1055         struct hlist_head *hash_list = &cfqd->cfq_hash[hashval];
1056         struct hlist_node *entry;
1057         struct cfq_queue *__cfqq;
1058
1059         hlist_for_each_entry(__cfqq, entry, hash_list, cfq_hash) {
1060                 const unsigned short __p = IOPRIO_PRIO_VALUE(__cfqq->org_ioprio_class, __cfqq->org_ioprio);
1061
1062                 if (__cfqq->key == key && (__p == prio || !prio))
1063                         return __cfqq;
1064         }
1065
1066         return NULL;
1067 }
1068
1069 static struct cfq_queue *
1070 cfq_find_cfq_hash(struct cfq_data *cfqd, unsigned int key, unsigned short prio)
1071 {
1072         return __cfq_find_cfq_hash(cfqd, key, prio, hash_long(key, CFQ_QHASH_SHIFT));
1073 }
1074
1075 static void cfq_free_io_context(struct io_context *ioc)
1076 {
1077         struct cfq_io_context *__cic;
1078         struct rb_node *n;
1079         int freed = 0;
1080
1081         while ((n = rb_first(&ioc->cic_root)) != NULL) {
1082                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1083                 rb_erase(&__cic->rb_node, &ioc->cic_root);
1084                 kmem_cache_free(cfq_ioc_pool, __cic);
1085                 freed++;
1086         }
1087
1088         elv_ioc_count_mod(ioc_count, -freed);
1089
1090         if (ioc_gone && !elv_ioc_count_read(ioc_count))
1091                 complete(ioc_gone);
1092 }
1093
1094 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1095 {
1096         if (unlikely(cfqq == cfqd->active_queue))
1097                 __cfq_slice_expired(cfqd, cfqq, 0);
1098
1099         cfq_put_queue(cfqq);
1100 }
1101
1102 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1103                                          struct cfq_io_context *cic)
1104 {
1105         list_del_init(&cic->queue_list);
1106         smp_wmb();
1107         cic->key = NULL;
1108
1109         if (cic->cfqq[ASYNC]) {
1110                 cfq_exit_cfqq(cfqd, cic->cfqq[ASYNC]);
1111                 cic->cfqq[ASYNC] = NULL;
1112         }
1113
1114         if (cic->cfqq[SYNC]) {
1115                 cfq_exit_cfqq(cfqd, cic->cfqq[SYNC]);
1116                 cic->cfqq[SYNC] = NULL;
1117         }
1118 }
1119
1120
1121 /*
1122  * Called with interrupts disabled
1123  */
1124 static void cfq_exit_single_io_context(struct cfq_io_context *cic)
1125 {
1126         struct cfq_data *cfqd = cic->key;
1127
1128         if (cfqd) {
1129                 request_queue_t *q = cfqd->queue;
1130
1131                 spin_lock_irq(q->queue_lock);
1132                 __cfq_exit_single_io_context(cfqd, cic);
1133                 spin_unlock_irq(q->queue_lock);
1134         }
1135 }
1136
1137 static void cfq_exit_io_context(struct io_context *ioc)
1138 {
1139         struct cfq_io_context *__cic;
1140         struct rb_node *n;
1141
1142         /*
1143          * put the reference this task is holding to the various queues
1144          */
1145
1146         n = rb_first(&ioc->cic_root);
1147         while (n != NULL) {
1148                 __cic = rb_entry(n, struct cfq_io_context, rb_node);
1149
1150                 cfq_exit_single_io_context(__cic);
1151                 n = rb_next(n);
1152         }
1153 }
1154
1155 static struct cfq_io_context *
1156 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1157 {
1158         struct cfq_io_context *cic;
1159
1160         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask, cfqd->queue->node);
1161         if (cic) {
1162                 memset(cic, 0, sizeof(*cic));
1163                 cic->last_end_request = jiffies;
1164                 INIT_LIST_HEAD(&cic->queue_list);
1165                 cic->dtor = cfq_free_io_context;
1166                 cic->exit = cfq_exit_io_context;
1167                 elv_ioc_count_inc(ioc_count);
1168         }
1169
1170         return cic;
1171 }
1172
1173 static void cfq_init_prio_data(struct cfq_queue *cfqq)
1174 {
1175         struct task_struct *tsk = current;
1176         int ioprio_class;
1177
1178         if (!cfq_cfqq_prio_changed(cfqq))
1179                 return;
1180
1181         ioprio_class = IOPRIO_PRIO_CLASS(tsk->ioprio);
1182         switch (ioprio_class) {
1183                 default:
1184                         printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1185                 case IOPRIO_CLASS_NONE:
1186                         /*
1187                          * no prio set, place us in the middle of the BE classes
1188                          */
1189                         cfqq->ioprio = task_nice_ioprio(tsk);
1190                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1191                         break;
1192                 case IOPRIO_CLASS_RT:
1193                         cfqq->ioprio = task_ioprio(tsk);
1194                         cfqq->ioprio_class = IOPRIO_CLASS_RT;
1195                         break;
1196                 case IOPRIO_CLASS_BE:
1197                         cfqq->ioprio = task_ioprio(tsk);
1198                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1199                         break;
1200                 case IOPRIO_CLASS_IDLE:
1201                         cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1202                         cfqq->ioprio = 7;
1203                         cfq_clear_cfqq_idle_window(cfqq);
1204                         break;
1205         }
1206
1207         /*
1208          * keep track of original prio settings in case we have to temporarily
1209          * elevate the priority of this queue
1210          */
1211         cfqq->org_ioprio = cfqq->ioprio;
1212         cfqq->org_ioprio_class = cfqq->ioprio_class;
1213
1214         if (cfq_cfqq_on_rr(cfqq))
1215                 cfq_resort_rr_list(cfqq, 0);
1216
1217         cfq_clear_cfqq_prio_changed(cfqq);
1218 }
1219
1220 static inline void changed_ioprio(struct cfq_io_context *cic)
1221 {
1222         struct cfq_data *cfqd = cic->key;
1223         struct cfq_queue *cfqq;
1224         unsigned long flags;
1225
1226         if (unlikely(!cfqd))
1227                 return;
1228
1229         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1230
1231         cfqq = cic->cfqq[ASYNC];
1232         if (cfqq) {
1233                 struct cfq_queue *new_cfqq;
1234                 new_cfqq = cfq_get_queue(cfqd, CFQ_KEY_ASYNC, cic->ioc->task,
1235                                          GFP_ATOMIC);
1236                 if (new_cfqq) {
1237                         cic->cfqq[ASYNC] = new_cfqq;
1238                         cfq_put_queue(cfqq);
1239                 }
1240         }
1241
1242         cfqq = cic->cfqq[SYNC];
1243         if (cfqq)
1244                 cfq_mark_cfqq_prio_changed(cfqq);
1245
1246         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1247 }
1248
1249 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1250 {
1251         struct cfq_io_context *cic;
1252         struct rb_node *n;
1253
1254         ioc->ioprio_changed = 0;
1255
1256         n = rb_first(&ioc->cic_root);
1257         while (n != NULL) {
1258                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1259
1260                 changed_ioprio(cic);
1261                 n = rb_next(n);
1262         }
1263 }
1264
1265 static struct cfq_queue *
1266 cfq_get_queue(struct cfq_data *cfqd, unsigned int key, struct task_struct *tsk,
1267               gfp_t gfp_mask)
1268 {
1269         const int hashval = hash_long(key, CFQ_QHASH_SHIFT);
1270         struct cfq_queue *cfqq, *new_cfqq = NULL;
1271         unsigned short ioprio;
1272
1273 retry:
1274         ioprio = tsk->ioprio;
1275         cfqq = __cfq_find_cfq_hash(cfqd, key, ioprio, hashval);
1276
1277         if (!cfqq) {
1278                 if (new_cfqq) {
1279                         cfqq = new_cfqq;
1280                         new_cfqq = NULL;
1281                 } else if (gfp_mask & __GFP_WAIT) {
1282                         /*
1283                          * Inform the allocator of the fact that we will
1284                          * just repeat this allocation if it fails, to allow
1285                          * the allocator to do whatever it needs to attempt to
1286                          * free memory.
1287                          */
1288                         spin_unlock_irq(cfqd->queue->queue_lock);
1289                         new_cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask|__GFP_NOFAIL, cfqd->queue->node);
1290                         spin_lock_irq(cfqd->queue->queue_lock);
1291                         goto retry;
1292                 } else {
1293                         cfqq = kmem_cache_alloc_node(cfq_pool, gfp_mask, cfqd->queue->node);
1294                         if (!cfqq)
1295                                 goto out;
1296                 }
1297
1298                 memset(cfqq, 0, sizeof(*cfqq));
1299
1300                 INIT_HLIST_NODE(&cfqq->cfq_hash);
1301                 INIT_LIST_HEAD(&cfqq->cfq_list);
1302                 INIT_LIST_HEAD(&cfqq->fifo);
1303
1304                 cfqq->key = key;
1305                 hlist_add_head(&cfqq->cfq_hash, &cfqd->cfq_hash[hashval]);
1306                 atomic_set(&cfqq->ref, 0);
1307                 cfqq->cfqd = cfqd;
1308                 /*
1309                  * set ->slice_left to allow preemption for a new process
1310                  */
1311                 cfqq->slice_left = 2 * cfqd->cfq_slice_idle;
1312                 cfq_mark_cfqq_idle_window(cfqq);
1313                 cfq_mark_cfqq_prio_changed(cfqq);
1314                 cfq_mark_cfqq_queue_new(cfqq);
1315                 cfq_init_prio_data(cfqq);
1316         }
1317
1318         if (new_cfqq)
1319                 kmem_cache_free(cfq_pool, new_cfqq);
1320
1321         atomic_inc(&cfqq->ref);
1322 out:
1323         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1324         return cfqq;
1325 }
1326
1327 static void
1328 cfq_drop_dead_cic(struct io_context *ioc, struct cfq_io_context *cic)
1329 {
1330         WARN_ON(!list_empty(&cic->queue_list));
1331         rb_erase(&cic->rb_node, &ioc->cic_root);
1332         kmem_cache_free(cfq_ioc_pool, cic);
1333         elv_ioc_count_dec(ioc_count);
1334 }
1335
1336 static struct cfq_io_context *
1337 cfq_cic_rb_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1338 {
1339         struct rb_node *n;
1340         struct cfq_io_context *cic;
1341         void *k, *key = cfqd;
1342
1343 restart:
1344         n = ioc->cic_root.rb_node;
1345         while (n) {
1346                 cic = rb_entry(n, struct cfq_io_context, rb_node);
1347                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1348                 k = cic->key;
1349                 if (unlikely(!k)) {
1350                         cfq_drop_dead_cic(ioc, cic);
1351                         goto restart;
1352                 }
1353
1354                 if (key < k)
1355                         n = n->rb_left;
1356                 else if (key > k)
1357                         n = n->rb_right;
1358                 else
1359                         return cic;
1360         }
1361
1362         return NULL;
1363 }
1364
1365 static inline void
1366 cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1367              struct cfq_io_context *cic)
1368 {
1369         struct rb_node **p;
1370         struct rb_node *parent;
1371         struct cfq_io_context *__cic;
1372         unsigned long flags;
1373         void *k;
1374
1375         cic->ioc = ioc;
1376         cic->key = cfqd;
1377
1378 restart:
1379         parent = NULL;
1380         p = &ioc->cic_root.rb_node;
1381         while (*p) {
1382                 parent = *p;
1383                 __cic = rb_entry(parent, struct cfq_io_context, rb_node);
1384                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1385                 k = __cic->key;
1386                 if (unlikely(!k)) {
1387                         cfq_drop_dead_cic(ioc, __cic);
1388                         goto restart;
1389                 }
1390
1391                 if (cic->key < k)
1392                         p = &(*p)->rb_left;
1393                 else if (cic->key > k)
1394                         p = &(*p)->rb_right;
1395                 else
1396                         BUG();
1397         }
1398
1399         rb_link_node(&cic->rb_node, parent, p);
1400         rb_insert_color(&cic->rb_node, &ioc->cic_root);
1401
1402         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1403         list_add(&cic->queue_list, &cfqd->cic_list);
1404         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1405 }
1406
1407 /*
1408  * Setup general io context and cfq io context. There can be several cfq
1409  * io contexts per general io context, if this process is doing io to more
1410  * than one device managed by cfq.
1411  */
1412 static struct cfq_io_context *
1413 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1414 {
1415         struct io_context *ioc = NULL;
1416         struct cfq_io_context *cic;
1417
1418         might_sleep_if(gfp_mask & __GFP_WAIT);
1419
1420         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1421         if (!ioc)
1422                 return NULL;
1423
1424         cic = cfq_cic_rb_lookup(cfqd, ioc);
1425         if (cic)
1426                 goto out;
1427
1428         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1429         if (cic == NULL)
1430                 goto err;
1431
1432         cfq_cic_link(cfqd, ioc, cic);
1433 out:
1434         smp_read_barrier_depends();
1435         if (unlikely(ioc->ioprio_changed))
1436                 cfq_ioc_set_ioprio(ioc);
1437
1438         return cic;
1439 err:
1440         put_io_context(ioc);
1441         return NULL;
1442 }
1443
1444 static void
1445 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1446 {
1447         unsigned long elapsed, ttime;
1448
1449         /*
1450          * if this context already has stuff queued, thinktime is from
1451          * last queue not last end
1452          */
1453 #if 0
1454         if (time_after(cic->last_end_request, cic->last_queue))
1455                 elapsed = jiffies - cic->last_end_request;
1456         else
1457                 elapsed = jiffies - cic->last_queue;
1458 #else
1459                 elapsed = jiffies - cic->last_end_request;
1460 #endif
1461
1462         ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1463
1464         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1465         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1466         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1467 }
1468
1469 static void
1470 cfq_update_io_seektime(struct cfq_io_context *cic, struct request *rq)
1471 {
1472         sector_t sdist;
1473         u64 total;
1474
1475         if (cic->last_request_pos < rq->sector)
1476                 sdist = rq->sector - cic->last_request_pos;
1477         else
1478                 sdist = cic->last_request_pos - rq->sector;
1479
1480         /*
1481          * Don't allow the seek distance to get too large from the
1482          * odd fragment, pagein, etc
1483          */
1484         if (cic->seek_samples <= 60) /* second&third seek */
1485                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1486         else
1487                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1488
1489         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1490         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1491         total = cic->seek_total + (cic->seek_samples/2);
1492         do_div(total, cic->seek_samples);
1493         cic->seek_mean = (sector_t)total;
1494 }
1495
1496 /*
1497  * Disable idle window if the process thinks too long or seeks so much that
1498  * it doesn't matter
1499  */
1500 static void
1501 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1502                        struct cfq_io_context *cic)
1503 {
1504         int enable_idle = cfq_cfqq_idle_window(cfqq);
1505
1506         if (!cic->ioc->task || !cfqd->cfq_slice_idle ||
1507             (cfqd->hw_tag && CIC_SEEKY(cic)))
1508                 enable_idle = 0;
1509         else if (sample_valid(cic->ttime_samples)) {
1510                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1511                         enable_idle = 0;
1512                 else
1513                         enable_idle = 1;
1514         }
1515
1516         if (enable_idle)
1517                 cfq_mark_cfqq_idle_window(cfqq);
1518         else
1519                 cfq_clear_cfqq_idle_window(cfqq);
1520 }
1521
1522
1523 /*
1524  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1525  * no or if we aren't sure, a 1 will cause a preempt.
1526  */
1527 static int
1528 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1529                    struct request *rq)
1530 {
1531         struct cfq_queue *cfqq = cfqd->active_queue;
1532
1533         if (cfq_class_idle(new_cfqq))
1534                 return 0;
1535
1536         if (!cfqq)
1537                 return 0;
1538
1539         if (cfq_class_idle(cfqq))
1540                 return 1;
1541         if (!cfq_cfqq_wait_request(new_cfqq))
1542                 return 0;
1543         /*
1544          * if it doesn't have slice left, forget it
1545          */
1546         if (new_cfqq->slice_left < cfqd->cfq_slice_idle)
1547                 return 0;
1548         /*
1549          * if the new request is sync, but the currently running queue is
1550          * not, let the sync request have priority.
1551          */
1552         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1553                 return 1;
1554         /*
1555          * So both queues are sync. Let the new request get disk time if
1556          * it's a metadata request and the current queue is doing regular IO.
1557          */
1558         if (rq_is_meta(rq) && !cfqq->meta_pending)
1559                 return 1;
1560
1561         return 0;
1562 }
1563
1564 /*
1565  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1566  * let it have half of its nominal slice.
1567  */
1568 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1569 {
1570         cfq_slice_expired(cfqd, 1);
1571
1572         if (!cfqq->slice_left)
1573                 cfqq->slice_left = cfq_prio_to_slice(cfqd, cfqq) / 2;
1574
1575         /*
1576          * Put the new queue at the front of the of the current list,
1577          * so we know that it will be selected next.
1578          */
1579         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1580         list_move(&cfqq->cfq_list, &cfqd->cur_rr);
1581
1582         cfqq->slice_end = cfqq->slice_left + jiffies;
1583 }
1584
1585 /*
1586  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1587  * something we should do about it
1588  */
1589 static void
1590 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1591                 struct request *rq)
1592 {
1593         struct cfq_io_context *cic = RQ_CIC(rq);
1594
1595         if (rq_is_meta(rq))
1596                 cfqq->meta_pending++;
1597
1598         /*
1599          * check if this request is a better next-serve candidate)) {
1600          */
1601         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
1602         BUG_ON(!cfqq->next_rq);
1603
1604         /*
1605          * we never wait for an async request and we don't allow preemption
1606          * of an async request. so just return early
1607          */
1608         if (!rq_is_sync(rq)) {
1609                 /*
1610                  * sync process issued an async request, if it's waiting
1611                  * then expire it and kick rq handling.
1612                  */
1613                 if (cic == cfqd->active_cic &&
1614                     del_timer(&cfqd->idle_slice_timer)) {
1615                         cfq_slice_expired(cfqd, 0);
1616                         blk_start_queueing(cfqd->queue);
1617                 }
1618                 return;
1619         }
1620
1621         cfq_update_io_thinktime(cfqd, cic);
1622         cfq_update_io_seektime(cic, rq);
1623         cfq_update_idle_window(cfqd, cfqq, cic);
1624
1625         cic->last_queue = jiffies;
1626         cic->last_request_pos = rq->sector + rq->nr_sectors;
1627
1628         if (cfqq == cfqd->active_queue) {
1629                 /*
1630                  * if we are waiting for a request for this queue, let it rip
1631                  * immediately and flag that we must not expire this queue
1632                  * just now
1633                  */
1634                 if (cfq_cfqq_wait_request(cfqq)) {
1635                         cfq_mark_cfqq_must_dispatch(cfqq);
1636                         del_timer(&cfqd->idle_slice_timer);
1637                         blk_start_queueing(cfqd->queue);
1638                 }
1639         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1640                 /*
1641                  * not the active queue - expire current slice if it is
1642                  * idle and has expired it's mean thinktime or this new queue
1643                  * has some old slice time left and is of higher priority
1644                  */
1645                 cfq_preempt_queue(cfqd, cfqq);
1646                 cfq_mark_cfqq_must_dispatch(cfqq);
1647                 blk_start_queueing(cfqd->queue);
1648         }
1649 }
1650
1651 static void cfq_insert_request(request_queue_t *q, struct request *rq)
1652 {
1653         struct cfq_data *cfqd = q->elevator->elevator_data;
1654         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1655
1656         cfq_init_prio_data(cfqq);
1657
1658         cfq_add_rq_rb(rq);
1659
1660         list_add_tail(&rq->queuelist, &cfqq->fifo);
1661
1662         cfq_rq_enqueued(cfqd, cfqq, rq);
1663 }
1664
1665 static void cfq_completed_request(request_queue_t *q, struct request *rq)
1666 {
1667         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1668         struct cfq_data *cfqd = cfqq->cfqd;
1669         const int sync = rq_is_sync(rq);
1670         unsigned long now;
1671
1672         now = jiffies;
1673
1674         WARN_ON(!cfqd->rq_in_driver);
1675         WARN_ON(!cfqq->on_dispatch[sync]);
1676         cfqd->rq_in_driver--;
1677         cfqq->on_dispatch[sync]--;
1678
1679         if (!cfq_class_idle(cfqq))
1680                 cfqd->last_end_request = now;
1681
1682         if (!cfq_cfqq_dispatched(cfqq) && cfq_cfqq_on_rr(cfqq))
1683                 cfq_resort_rr_list(cfqq, 0);
1684
1685         if (sync)
1686                 RQ_CIC(rq)->last_end_request = now;
1687
1688         /*
1689          * If this is the active queue, check if it needs to be expired,
1690          * or if we want to idle in case it has no pending requests.
1691          */
1692         if (cfqd->active_queue == cfqq) {
1693                 if (time_after(now, cfqq->slice_end))
1694                         cfq_slice_expired(cfqd, 0);
1695                 else if (sync && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1696                         if (!cfq_arm_slice_timer(cfqd, cfqq))
1697                                 cfq_schedule_dispatch(cfqd);
1698                 }
1699         }
1700 }
1701
1702 /*
1703  * we temporarily boost lower priority queues if they are holding fs exclusive
1704  * resources. they are boosted to normal prio (CLASS_BE/4)
1705  */
1706 static void cfq_prio_boost(struct cfq_queue *cfqq)
1707 {
1708         const int ioprio_class = cfqq->ioprio_class;
1709         const int ioprio = cfqq->ioprio;
1710
1711         if (has_fs_excl()) {
1712                 /*
1713                  * boost idle prio on transactions that would lock out other
1714                  * users of the filesystem
1715                  */
1716                 if (cfq_class_idle(cfqq))
1717                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
1718                 if (cfqq->ioprio > IOPRIO_NORM)
1719                         cfqq->ioprio = IOPRIO_NORM;
1720         } else {
1721                 /*
1722                  * check if we need to unboost the queue
1723                  */
1724                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
1725                         cfqq->ioprio_class = cfqq->org_ioprio_class;
1726                 if (cfqq->ioprio != cfqq->org_ioprio)
1727                         cfqq->ioprio = cfqq->org_ioprio;
1728         }
1729
1730         /*
1731          * refile between round-robin lists if we moved the priority class
1732          */
1733         if ((ioprio_class != cfqq->ioprio_class || ioprio != cfqq->ioprio) &&
1734             cfq_cfqq_on_rr(cfqq))
1735                 cfq_resort_rr_list(cfqq, 0);
1736 }
1737
1738 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
1739 {
1740         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
1741             !cfq_cfqq_must_alloc_slice(cfqq)) {
1742                 cfq_mark_cfqq_must_alloc_slice(cfqq);
1743                 return ELV_MQUEUE_MUST;
1744         }
1745
1746         return ELV_MQUEUE_MAY;
1747 }
1748
1749 static int cfq_may_queue(request_queue_t *q, int rw)
1750 {
1751         struct cfq_data *cfqd = q->elevator->elevator_data;
1752         struct task_struct *tsk = current;
1753         struct cfq_queue *cfqq;
1754         unsigned int key;
1755
1756         key = cfq_queue_pid(tsk, rw, rw & REQ_RW_SYNC);
1757
1758         /*
1759          * don't force setup of a queue from here, as a call to may_queue
1760          * does not necessarily imply that a request actually will be queued.
1761          * so just lookup a possibly existing queue, or return 'may queue'
1762          * if that fails
1763          */
1764         cfqq = cfq_find_cfq_hash(cfqd, key, tsk->ioprio);
1765         if (cfqq) {
1766                 cfq_init_prio_data(cfqq);
1767                 cfq_prio_boost(cfqq);
1768
1769                 return __cfq_may_queue(cfqq);
1770         }
1771
1772         return ELV_MQUEUE_MAY;
1773 }
1774
1775 /*
1776  * queue lock held here
1777  */
1778 static void cfq_put_request(struct request *rq)
1779 {
1780         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1781
1782         if (cfqq) {
1783                 const int rw = rq_data_dir(rq);
1784
1785                 BUG_ON(!cfqq->allocated[rw]);
1786                 cfqq->allocated[rw]--;
1787
1788                 put_io_context(RQ_CIC(rq)->ioc);
1789
1790                 rq->elevator_private = NULL;
1791                 rq->elevator_private2 = NULL;
1792
1793                 cfq_put_queue(cfqq);
1794         }
1795 }
1796
1797 /*
1798  * Allocate cfq data structures associated with this request.
1799  */
1800 static int
1801 cfq_set_request(request_queue_t *q, struct request *rq, gfp_t gfp_mask)
1802 {
1803         struct cfq_data *cfqd = q->elevator->elevator_data;
1804         struct task_struct *tsk = current;
1805         struct cfq_io_context *cic;
1806         const int rw = rq_data_dir(rq);
1807         const int is_sync = rq_is_sync(rq);
1808         pid_t key = cfq_queue_pid(tsk, rw, is_sync);
1809         struct cfq_queue *cfqq;
1810         unsigned long flags;
1811
1812         might_sleep_if(gfp_mask & __GFP_WAIT);
1813
1814         cic = cfq_get_io_context(cfqd, gfp_mask);
1815
1816         spin_lock_irqsave(q->queue_lock, flags);
1817
1818         if (!cic)
1819                 goto queue_fail;
1820
1821         if (!cic->cfqq[is_sync]) {
1822                 cfqq = cfq_get_queue(cfqd, key, tsk, gfp_mask);
1823                 if (!cfqq)
1824                         goto queue_fail;
1825
1826                 cic->cfqq[is_sync] = cfqq;
1827         } else
1828                 cfqq = cic->cfqq[is_sync];
1829
1830         cfqq->allocated[rw]++;
1831         cfq_clear_cfqq_must_alloc(cfqq);
1832         atomic_inc(&cfqq->ref);
1833
1834         spin_unlock_irqrestore(q->queue_lock, flags);
1835
1836         rq->elevator_private = cic;
1837         rq->elevator_private2 = cfqq;
1838         return 0;
1839
1840 queue_fail:
1841         if (cic)
1842                 put_io_context(cic->ioc);
1843
1844         cfq_schedule_dispatch(cfqd);
1845         spin_unlock_irqrestore(q->queue_lock, flags);
1846         return 1;
1847 }
1848
1849 static void cfq_kick_queue(struct work_struct *work)
1850 {
1851         struct cfq_data *cfqd =
1852                 container_of(work, struct cfq_data, unplug_work);
1853         request_queue_t *q = cfqd->queue;
1854         unsigned long flags;
1855
1856         spin_lock_irqsave(q->queue_lock, flags);
1857         blk_start_queueing(q);
1858         spin_unlock_irqrestore(q->queue_lock, flags);
1859 }
1860
1861 /*
1862  * Timer running if the active_queue is currently idling inside its time slice
1863  */
1864 static void cfq_idle_slice_timer(unsigned long data)
1865 {
1866         struct cfq_data *cfqd = (struct cfq_data *) data;
1867         struct cfq_queue *cfqq;
1868         unsigned long flags;
1869
1870         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1871
1872         if ((cfqq = cfqd->active_queue) != NULL) {
1873                 unsigned long now = jiffies;
1874
1875                 /*
1876                  * expired
1877                  */
1878                 if (time_after(now, cfqq->slice_end))
1879                         goto expire;
1880
1881                 /*
1882                  * only expire and reinvoke request handler, if there are
1883                  * other queues with pending requests
1884                  */
1885                 if (!cfqd->busy_queues)
1886                         goto out_cont;
1887
1888                 /*
1889                  * not expired and it has a request pending, let it dispatch
1890                  */
1891                 if (!RB_EMPTY_ROOT(&cfqq->sort_list)) {
1892                         cfq_mark_cfqq_must_dispatch(cfqq);
1893                         goto out_kick;
1894                 }
1895         }
1896 expire:
1897         cfq_slice_expired(cfqd, 0);
1898 out_kick:
1899         cfq_schedule_dispatch(cfqd);
1900 out_cont:
1901         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1902 }
1903
1904 /*
1905  * Timer running if an idle class queue is waiting for service
1906  */
1907 static void cfq_idle_class_timer(unsigned long data)
1908 {
1909         struct cfq_data *cfqd = (struct cfq_data *) data;
1910         unsigned long flags, end;
1911
1912         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1913
1914         /*
1915          * race with a non-idle queue, reset timer
1916          */
1917         end = cfqd->last_end_request + CFQ_IDLE_GRACE;
1918         if (!time_after_eq(jiffies, end))
1919                 mod_timer(&cfqd->idle_class_timer, end);
1920         else
1921                 cfq_schedule_dispatch(cfqd);
1922
1923         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1924 }
1925
1926 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
1927 {
1928         del_timer_sync(&cfqd->idle_slice_timer);
1929         del_timer_sync(&cfqd->idle_class_timer);
1930         blk_sync_queue(cfqd->queue);
1931 }
1932
1933 static void cfq_exit_queue(elevator_t *e)
1934 {
1935         struct cfq_data *cfqd = e->elevator_data;
1936         request_queue_t *q = cfqd->queue;
1937
1938         cfq_shutdown_timer_wq(cfqd);
1939
1940         spin_lock_irq(q->queue_lock);
1941
1942         if (cfqd->active_queue)
1943                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
1944
1945         while (!list_empty(&cfqd->cic_list)) {
1946                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
1947                                                         struct cfq_io_context,
1948                                                         queue_list);
1949
1950                 __cfq_exit_single_io_context(cfqd, cic);
1951         }
1952
1953         spin_unlock_irq(q->queue_lock);
1954
1955         cfq_shutdown_timer_wq(cfqd);
1956
1957         kfree(cfqd->cfq_hash);
1958         kfree(cfqd);
1959 }
1960
1961 static void *cfq_init_queue(request_queue_t *q)
1962 {
1963         struct cfq_data *cfqd;
1964         int i;
1965
1966         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node);
1967         if (!cfqd)
1968                 return NULL;
1969
1970         memset(cfqd, 0, sizeof(*cfqd));
1971
1972         for (i = 0; i < CFQ_PRIO_LISTS; i++)
1973                 INIT_LIST_HEAD(&cfqd->rr_list[i]);
1974
1975         INIT_LIST_HEAD(&cfqd->busy_rr);
1976         INIT_LIST_HEAD(&cfqd->cur_rr);
1977         INIT_LIST_HEAD(&cfqd->idle_rr);
1978         INIT_LIST_HEAD(&cfqd->cic_list);
1979
1980         cfqd->cfq_hash = kmalloc_node(sizeof(struct hlist_head) * CFQ_QHASH_ENTRIES, GFP_KERNEL, q->node);
1981         if (!cfqd->cfq_hash)
1982                 goto out_free;
1983
1984         for (i = 0; i < CFQ_QHASH_ENTRIES; i++)
1985                 INIT_HLIST_HEAD(&cfqd->cfq_hash[i]);
1986
1987         cfqd->queue = q;
1988
1989         init_timer(&cfqd->idle_slice_timer);
1990         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
1991         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
1992
1993         init_timer(&cfqd->idle_class_timer);
1994         cfqd->idle_class_timer.function = cfq_idle_class_timer;
1995         cfqd->idle_class_timer.data = (unsigned long) cfqd;
1996
1997         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
1998
1999         cfqd->cfq_quantum = cfq_quantum;
2000         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2001         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2002         cfqd->cfq_back_max = cfq_back_max;
2003         cfqd->cfq_back_penalty = cfq_back_penalty;
2004         cfqd->cfq_slice[0] = cfq_slice_async;
2005         cfqd->cfq_slice[1] = cfq_slice_sync;
2006         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2007         cfqd->cfq_slice_idle = cfq_slice_idle;
2008
2009         return cfqd;
2010 out_free:
2011         kfree(cfqd);
2012         return NULL;
2013 }
2014
2015 static void cfq_slab_kill(void)
2016 {
2017         if (cfq_pool)
2018                 kmem_cache_destroy(cfq_pool);
2019         if (cfq_ioc_pool)
2020                 kmem_cache_destroy(cfq_ioc_pool);
2021 }
2022
2023 static int __init cfq_slab_setup(void)
2024 {
2025         cfq_pool = kmem_cache_create("cfq_pool", sizeof(struct cfq_queue), 0, 0,
2026                                         NULL, NULL);
2027         if (!cfq_pool)
2028                 goto fail;
2029
2030         cfq_ioc_pool = kmem_cache_create("cfq_ioc_pool",
2031                         sizeof(struct cfq_io_context), 0, 0, NULL, NULL);
2032         if (!cfq_ioc_pool)
2033                 goto fail;
2034
2035         return 0;
2036 fail:
2037         cfq_slab_kill();
2038         return -ENOMEM;
2039 }
2040
2041 /*
2042  * sysfs parts below -->
2043  */
2044
2045 static ssize_t
2046 cfq_var_show(unsigned int var, char *page)
2047 {
2048         return sprintf(page, "%d\n", var);
2049 }
2050
2051 static ssize_t
2052 cfq_var_store(unsigned int *var, const char *page, size_t count)
2053 {
2054         char *p = (char *) page;
2055
2056         *var = simple_strtoul(p, &p, 10);
2057         return count;
2058 }
2059
2060 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2061 static ssize_t __FUNC(elevator_t *e, char *page)                        \
2062 {                                                                       \
2063         struct cfq_data *cfqd = e->elevator_data;                       \
2064         unsigned int __data = __VAR;                                    \
2065         if (__CONV)                                                     \
2066                 __data = jiffies_to_msecs(__data);                      \
2067         return cfq_var_show(__data, (page));                            \
2068 }
2069 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2070 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2071 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2072 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2073 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2074 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2075 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2076 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2077 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2078 #undef SHOW_FUNCTION
2079
2080 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2081 static ssize_t __FUNC(elevator_t *e, const char *page, size_t count)    \
2082 {                                                                       \
2083         struct cfq_data *cfqd = e->elevator_data;                       \
2084         unsigned int __data;                                            \
2085         int ret = cfq_var_store(&__data, (page), count);                \
2086         if (__data < (MIN))                                             \
2087                 __data = (MIN);                                         \
2088         else if (__data > (MAX))                                        \
2089                 __data = (MAX);                                         \
2090         if (__CONV)                                                     \
2091                 *(__PTR) = msecs_to_jiffies(__data);                    \
2092         else                                                            \
2093                 *(__PTR) = __data;                                      \
2094         return ret;                                                     \
2095 }
2096 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2097 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, UINT_MAX, 1);
2098 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, UINT_MAX, 1);
2099 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2100 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, UINT_MAX, 0);
2101 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2102 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2103 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2104 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, UINT_MAX, 0);
2105 #undef STORE_FUNCTION
2106
2107 #define CFQ_ATTR(name) \
2108         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2109
2110 static struct elv_fs_entry cfq_attrs[] = {
2111         CFQ_ATTR(quantum),
2112         CFQ_ATTR(fifo_expire_sync),
2113         CFQ_ATTR(fifo_expire_async),
2114         CFQ_ATTR(back_seek_max),
2115         CFQ_ATTR(back_seek_penalty),
2116         CFQ_ATTR(slice_sync),
2117         CFQ_ATTR(slice_async),
2118         CFQ_ATTR(slice_async_rq),
2119         CFQ_ATTR(slice_idle),
2120         __ATTR_NULL
2121 };
2122
2123 static struct elevator_type iosched_cfq = {
2124         .ops = {
2125                 .elevator_merge_fn =            cfq_merge,
2126                 .elevator_merged_fn =           cfq_merged_request,
2127                 .elevator_merge_req_fn =        cfq_merged_requests,
2128                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2129                 .elevator_add_req_fn =          cfq_insert_request,
2130                 .elevator_activate_req_fn =     cfq_activate_request,
2131                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2132                 .elevator_queue_empty_fn =      cfq_queue_empty,
2133                 .elevator_completed_req_fn =    cfq_completed_request,
2134                 .elevator_former_req_fn =       elv_rb_former_request,
2135                 .elevator_latter_req_fn =       elv_rb_latter_request,
2136                 .elevator_set_req_fn =          cfq_set_request,
2137                 .elevator_put_req_fn =          cfq_put_request,
2138                 .elevator_may_queue_fn =        cfq_may_queue,
2139                 .elevator_init_fn =             cfq_init_queue,
2140                 .elevator_exit_fn =             cfq_exit_queue,
2141                 .trim =                         cfq_free_io_context,
2142         },
2143         .elevator_attrs =       cfq_attrs,
2144         .elevator_name =        "cfq",
2145         .elevator_owner =       THIS_MODULE,
2146 };
2147
2148 static int __init cfq_init(void)
2149 {
2150         int ret;
2151
2152         /*
2153          * could be 0 on HZ < 1000 setups
2154          */
2155         if (!cfq_slice_async)
2156                 cfq_slice_async = 1;
2157         if (!cfq_slice_idle)
2158                 cfq_slice_idle = 1;
2159
2160         if (cfq_slab_setup())
2161                 return -ENOMEM;
2162
2163         ret = elv_register(&iosched_cfq);
2164         if (ret)
2165                 cfq_slab_kill();
2166
2167         return ret;
2168 }
2169
2170 static void __exit cfq_exit(void)
2171 {
2172         DECLARE_COMPLETION_ONSTACK(all_gone);
2173         elv_unregister(&iosched_cfq);
2174         ioc_gone = &all_gone;
2175         /* ioc_gone's update must be visible before reading ioc_count */
2176         smp_wmb();
2177         if (elv_ioc_count_read(ioc_count))
2178                 wait_for_completion(ioc_gone);
2179         synchronize_rcu();
2180         cfq_slab_kill();
2181 }
2182
2183 module_init(cfq_init);
2184 module_exit(cfq_exit);
2185
2186 MODULE_AUTHOR("Jens Axboe");
2187 MODULE_LICENSE("GPL");
2188 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");